Mixed-Precision Computing for Scientific Discovery: Formats, Co-Design, and Responsible Approximation
Year: 2026
Type: preprint
Abstract: Reduced and mixed precision have moved from a niche optimization to a central design axis in scientific computing and engineering, driven by energy constraints, heterogeneous accelerators, and the convergence of simulation and machine learning. This paper organizes the landscape around seven coupled themes---number formats, floating-point emulation, emerging architectures, hardware/software co-design, relation to other approximations, software design, and precision as a multilevel resource ---and, for each theme, synthesizes the state of the art, future directions, and open questions. We emphasize \emph{energy per trusted solution} as the core objective, and we frame \say{recklessly responsible} computing as a pragmatic doctrine: exploit low precision aggressively, but with systematic detection, escalation, and certification pathways.
Source: arXiv (Cornell University)
Authors: Agullo, Emmanuel, Anzt, Hartwig, Bauer, Daniel, Bindel, David, Buttari, Alfredo, Calotoiu, Alexandru, Carson, Erin Claire, D'Ambra, Pasqua, Daužickaitė, Ieva, Demmel, James W., Dongarra, Jack, Duff, Iain, Fasi, Massimiliano, Göddeke, Dominik, Graillat, Stef, Hunhold, Laslo, Iakymchuk, Roman, Jézéquel, Fabienne, Kohl, Nils, Köstler, Harald, Kružík, Jakub, Langou, Julien, Li, Xiaoye Sherry, Ltaief, Hatem, Luszczek, Piotr, Ma, Yuxin, Mary, Theo, Mikaitis, Mantas, Ootomo, Hiroyuki, Osei-Kuffuor, Daniel, Quintana-Ortí, Enrique S., Rüde, Ulrich, Scott, Jennifer, Shalf, John, Stals, Linda, Tamstorf, Rasmus, Turek, Stefan, Vacek, Petr, Vieublé, Bastien, Yokota, Rio
Language: English
Subfield: Computational Theory and Mathematics
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funders: European High Performance Computing Joint Undertaking, National Science Foundation, U.S. Department of Energy, Univerzita Karlova v Praze, European Commission, Agence Nationale de la Recherche, National Natural Science Foundation of China, Generalitat Valenciana, Office of Science, Engineering and Physical Sciences Research Council, Singapore Institute of Manufacturing Technology, Agencia Estatal de Investigación, Advanced Scientific Computing Research, Lawrence Livermore National Laboratory, CHIST-ERA
Awards: G1088636562, G1286236842, G1321919621, G1356433355, G1422035950, G1429959982, G14554683, Collaborative Research: Frameworks: Basic ALgebra LIbraries for Sustainable Technology with Interdisciplinary Collaboration (BALLISTIC), G1799333409, G1863003419, G2145270357, G2689834324, G3136028036, G3220251042, G3327601824, G3565621276, Methods and Algorithms for Exascale, G3669113783, G3857581125, High Performance Computing software and tools, G4094752214, G419881185, G42916606, G4438335082, G4501827968, G4576448283, G4683226447, Informing Future Numerical Standards by Determining Features of Non-Standard Mathematical Hardware, G498139845, G5005885202, G5015276450, G5061205453, G5133784819, G5195188938, G526230902, G5264466959, MPiK: Mixed Precision in Krylov Subspace Methods, G5685548091, G574795485, Center of Excellence for Exascale CFD, FOSTERING THE EUROPEAN ENERGY TRANSITION WITH EXASCALE, G6652958823, G704498496, G7080048889, G7915876455, G792975244, G8130583117, G8326222658, Floating-Point Transformer 4, Mixed precision algorithms for High Performance Computing, G8533620077, G895790057, G938317089, G991281326
OpenAlex ID: W7214936288